日本一わかりやすいStableDiffusion WebUI AUTOMATIC1111(ローカル版)のインストール方法と基本的な使い方

Summary

このビデオスクリプトでは、AIが生成するキャラクターイラストを作成する方法を紹介しています。特に人気の高い「stable diffusion web UI, Automatic 1111」のインストールと使い方について、2023年2月に関連する手順を解説しています。必要な環境設定からPythonやGitのインストール、モデルの選択とダウンロード、VAEの適用方法、そして実際にAIを使って絵を生成するプロンプトの書き方とテクニックまで詳しく説明されています。さまざまなスタイルと機能を持つ複数のモデルを紹介し、AIイラストの進化と新しいモデルの登場にも触れています。

Takeaways

• 💻 スクリプトはAI画像生成技術について説明しており、特に人気のある「stable diffusion web UI, Automatic 1111」のインストールと使用方法を紹介している。
• 📷 AIは手描きイラストとAI生成イラストを組み合わせてキャラクターのイラストを作成することができる。
• 🖼️ スクリプトでは、AIが生成するさまざまなスタイルのイラストを紹介しており、その中にはスケッチタッチイラストや画像のようなイラスト、ふわふわイラストが含まれている。
• 🛠️ インストールには比較的高スペックのコンピュータが必要なため、Windows 10以降のOS、NVIDIA製のグラフィックカード、4GB以上のビデオメモリが推奨されている。
• 🔍 PythonとGitのインストールが前提とされており、Pythonのバージョンは3.10.6が指定されている。
• 📂 インストールフォルダ名には全角文字を避け、DドライブなどCドライブ以外を推奨している。また、インストールフォルダへのショートカットを作成することが望ましい。
• 🔗 Gitを使用してStable Diffusion Web UIのリポジトリをクローンし、必要なモデルをダウンロードして使用する。
• 🎨 モデルは目的に合わせて選択可能で、アニメキャラクタや実写風イラストに長けたモデルが存在する。
• 🌐 VAE（Variational Auto Encoder）の適用がイラストの質感に影響を与えるとされ、スクリプトではStability AIのVAE 840,000を使用する例が紹介されている。
• 🔧 プロンプト（ Prompt）の書き方と設定変更がAIが生成するイラストに大きな影響を与えるとされ、スタイル、環境、主体部を含めたプロンプトの組成方法が説明されている。
• 🎭 スクリプトでは、AIが提供する様々な機能についても触れており、テキストから画像生成（Text to Image）、画像から画像（Image to Image）、アップスケール、PNG情報の表示などが紹介されている。

Q & A

AIが生成するイラストの半分はどのように作られていると説明されていますか？

AIが生成するイラストの半分は手書きのイラストに基づいており、もう半分はAIによって生成されたイラストです。

Live2Dとウェブカメラを使って何ができるようになりますか？

Live2Dとウェブカメラを使って、絵を描くスキルがなくても自分好みのキャラクターのイラストを作成することができるようになります。

「stable diffusion web UI, Automatic 1111」は何で使われていると説明されていますか？

「stable diffusion web UI, Automatic 1111」はローカル環境で構築され、動作する最も人気のある画像生成AIです。

インストールに必要なPCのスペックはどの程度必要ですか？

インストールに必要なPCスペックは比較的高く、Windows 10以降のOSが推奨され、NVIDIA製のグラフィックカードと少なくとも4GBのビデオメモリが必要なとされています。

Pythonのバージョンについてどのような注意が必要ですか？

Pythonのバージョンは3.10.6を使用することが推奨されており、異なるバージョンではエラーが発生する可能性があると注意が必要です。

Stable Diffusion Web UIをインストールする際に、どのドライブにインストールするのが良いと提案されていますか？

Stable Diffusion Web UIはDドライブなどのCドライブ以外にインストールすることが推奨されており、SSDにインストールすることが好ましいと提案されています。

モデルファイルの種類についてどのように選ぶべきですか？

モデルファイルはprunedとsafetensorsバージョンがバランスが良く、セキュリティと容量の両方において優れているとされています。

VAEとは何で、どのような効果がありますか？

VAEは「Variational Auto Encoder」の略で、AIイラストにおいては、イラストの仕上がりを微妙に変化させる効果があります。

プロンプトの書き方にはどのようなルールがありますか？

プロンプトは前方の言葉が強く考慮され、スタイル、環境、メインボディーの3つの要素で品質を上げることが推奨されています。

エミュレーションの方法として何が挙げられていますか？

エミュレーションの方法として「Hypernetwork」、「Dreambooth」、「LoRA」などが挙げられています。

AIが生成するイラストで、人体が正常に表示されない場合、どうすればよいですか？

人体が正常に表示されない場合は、「Restore face」機能をオフにし、解像度やアスペクト比を調整することで改善が可能な場合があります。

Outlines

00:00

🧑‍💻 AI画像生成技術の紹介

この段落では、AI画像生成技術が最近のホットトピックであることと、特に注目されている「stable diffusion web UI, Automatic 1111」について紹介しています。AIを使って自分の好きなキャラクターのイラストを簡単に作成できる方法を説明しており、必要なPCスペックやOS要件、PythonやGitなどのインストール手順について詳述されています。また、Windows 10以降での操作方法や、ファイル拡張子の確認方法も説明されています。

05:01

📂 Stable Diffusion Web UIのインストール手順

段落2では、Stable Diffusion Web UIをインストールする具体的な手順が紹介されています。まず、PythonとGitのインストール後に、インストール先のフォルダを作成し、コマンドプロンプトからGitを使用してリポジトリをクローンする方法について説明されています。次に、モデルの選択とダウンロード、モデルファイルを配置する場所について詳しく説明しており、最初の起動時に必要な手順も紹介されています。

10:03

🖼️ AI画像生成の基本操作

この段落では、Stable Diffusion Web UIの基本的な操作方法が紹介されています。ウェブUIの起動方法と、最初の画像生成までの手順について説明されています。また、UI上部のタブ機能や各タブでの操作内容についても詳述しており、テキストから画像を生成する「Text to Image」タブの使い方など、基本的な画像生成プロセスを紹介しています。

15:05

🎨 プロンプトの書き方とVAEの適用

段落4では、AIが理解するプロンプトの書き方と、画像の質感を調整するVAE（Variational Auto Encoder）の適用方法について説明されています。プロンプトの書き方として、スタイル、環境、メインボディーの3つの要素を組み合わせる方法や、品質を向上させるための「品質スペル」について紹介されています。VAEの種類と、それを適用することで得られる効果についても詳述されています。

20:09

🔍 画像生成の詳細設定と強調表現

この段落では、画像生成の詳細設定と強調表現について紹介されています。生成方法の選択肢としてカーラス系列の方法が挙げられ、ステップ数やリゾリューションなどの設定項目について説明されています。また、強調表現として、ブラケットを使用して特定の属性を強調する方法や、その効果についても紹介されています。

25:12

🌟 おすすめAIモデルの紹介

最後の段落では、様々なAIモデルについて紹介されています。異なるモデルが生成する画像の質感の違いや、それぞれが得意とするイラストのスタイルについて説明されています。また、モデルの選択肢として、アニメ風から実写風まで幅広いモデルが紹介されており、AIイラスト技術の進化と新しい機能のリリースについても触れています。

Keywords

💡AI

AIとは「人工知能」の略で、コンピュータが人間のように思考し、判断し、行動する技術を指します。このビデオでは、AIを用いてキャラクターのイラストを生成する技術に焦点が当てられており、AIによって手書きのイラストと組み合わせて新しいアート作品が創造されています。

💡Live2D

Live2Dは、2Dのイラストを3Dのように動かすことができる技術です。ビデオでは、Live2Dを使ってキャラクターの動きを表現することができると説明されています。これは、静的なイラストから動的な表現への一歩として、創造的なプロセスに重要な役割を果たしています。

💡stable diffusion web UI, Automatic 1111

stable diffusion web UI, Automatic 1111は、ビデオの中心となるAI画像生成ソフトウェアです。ローカル環境で動作し、ユーザーが自分の好みに合わせてキャラクターイラストを生成できるツールです。ビデオでは、このソフトウェアのインストールと使用方法が詳しく紹介されています。

💡Python

Pythonは、普及しているプログラミング言語の一つです。ビデオでは、PythonのインストールがAIソフトウェアを実行するための前提条件として挙げられています。Pythonは、AI開発においても重要な言語であり、様々なAIアルゴリズムを実装するのに使われます。

💡Git

Gitは、ソフトウェア開発においてバージョン管理を行うツールです。ビデオでは、GitのインストールがAIソフトウェアのセットアップ手順の一部として必要とされています。Gitを通じて、ソフトウェアの開発者たちは共同でプロジェクトを進め、変更履歴を管理することができます。

💡NVIDIA graphics card

NVIDIA製のグラフィックカードは、ビデオでAIソフトウェアを実行するために推奨されるハードウェアの一つです。グラフィックカードは、画像処理能力が強く、AIの学習や画像生成プロセスで高速な処理が必要な時に特に重要です。

💡SSD

SSDとは「ソリッドステートドライブ」の略で、コンピュータのストレージデバイスです。ビデオでは、SSDへのソフトウェアのインストールが推奨されており、これにより起動や読み込みが高速化されると説明されています。SSDは、データの読み書き速度がHDDよりもはるかに早く、パフォーマンス向上に寄与しています。

💡Hugging Face

Hugging Faceは、AIモデルを共有・配布するプラットフォームです。ビデオでは、Hugging Faceから様々なAIモデルをダウンロードし、Stable Diffusion Web UIで使用することができると紹介されています。このプラットフォームは、AIコミュニティにとって重要なリソースです。

💡VAE

VAEとは「Variational Auto Encoder」の略で、AIで使われるアルゴリズムの一つです。ビデオでは、VAEを適用することでイラストの質感を変化させる方法が紹介されています。VAEは、生成されたアートワークの多様性や表現力を高めるために用いられます。

💡Text to Image

Text to Imageは、テキストを元に画像を生成するAIの機能です。ビデオでは、Text to Imageタブを使ってテキストを入力し、AIがそれをもとにイラストを生成するプロセスが説明されています。この機能は、創造性と柔軟性を持ち、ユーザーが自分の想像を視覚化するのに役立ちます。

💡PNG Info

PNG Infoは、生成されたPNG画像に含まれる情報を見ることができる機能です。ビデオでは、PNG Infoタブを使って過去の生成設定を確認し、再利用することができると紹介されています。この機能は、同じ結果を再現するのにも役立ちます。

Highlights

AI生成插图技术日益普及，可以创建手绘和AI生成图像的混合体。

介绍了如何安装和使用流行的本地环境图像生成AI 'stable diffusion web UI, Automatic 1111'。

安装AI需要较高的电脑配置，推荐使用Windows 10或更高版本的操作系统。

需要NVIDIA显卡和至少4GB的视频内存，理想情况下使用12GB以及SSD硬盘。

展示如何检查操作系统版本和文件扩展名显示设置。

详细步骤教导如何安装Python和Git，Stable Diffusion Web UI的前提条件。

解释了如何通过Git克隆Stable Diffusion Web UI的代码到本地文件夹。

强调了安装文件夹名称应使用单字节字符以避免安装错误。

介绍了如何获取和安装AI模型文件，例如动漫风格的AnythingV4模型。

演示了Stable Diffusion Web UI的首次启动过程及其界面概览。

解释了如何结束Web UI会话以及下次启动的方法。

介绍了“Xformers”选项，可以提高图像生成速度。

界面介绍包括Text to Image、Image to Image、Inpaint、Extra和PNG Info等标签页的功能。

高级用户功能包括Checkpoint Merger和Train标签页，用于模型合成和学习。

详细说明了Text to Image标签页的基本操作和如何生成图像。

解释了VAE（变分自编码器）的作用以及如何安装和应用VAE文件以改善图像质量。

讨论了编写提示（prompts）的技巧，包括质量、风格、环境和主体的组合。

介绍了如何使用强调提示（emphasized prompts）来控制AI生成图像的特定特征。

分享了多个作者推荐的不同风格的AI模型，如Anything V3、ACertainthing、Seventh Anime V3等。

强调AI插图技术正在快速发展，不断有新模型和功能被发布。

Transcripts

What will come out?

Soooo Cute!

Hello, I'm Robin.

I'm Teruru.

Recently, I've been hearing the word AI almost every day.

AI is everything in the world.

In fact, half of our bodies are hand-drawn illustrations, and the other half is based on AI-generated illustrations.

And motioned with Live2D and a web camera.

Even if you don't have any drawing skills, you can make your own favorite character illustrations.

This is the image-generating AI that is now a hot topic.

For example, such a sketch-touch illustration and a picture-like illustration are all generated by AI.

And you can make such a fluffy illustration.

This time, among the many image-generating AI,

I would like to introduce how to install and use the most popular "stable diffusion web UI, Automatic 1111", which is built and runs in a local environment.

It's a way of using it in February of 2023, so please understand that it may change in the future.

First of all, what you need to install.

Of course computer!

However, the required specs are quite high, and the OS is after Windows 10.

It seems that it can be built in Win7, Mac, and Linux environments, but it seems that many of the expansion functions are made in the 10or11 environment, so I recommend Windows after 10.

What do you do when you don't know the OS?

Press Windows and R keys at the same time on the desktop.

Select the file name and enter DXDIAG to see the dialog called execution.

Then the DirectX diagnosis tool will start, so you can see it in the operating system section in the system tab on page 1.

In this video, we will talk about the Windows 11 64-bit version.

And you need to have an NVIDIA-made graphics card.

In addition, it is difficult if you do not have at least 4GB of video memory.

If you just generate an illustration, it will work with 4GB, but it is ideal to have about 12GB when using the learning function.

I think it's better to have about 30GB of drive space.

This is to eat a lot of capacity, with a capacity of about 5GB per model.

Also, if it is a hard disk, it will be too late and your head will be bald, so I recommend installing it on SSD.

The last thing you should not forget is to display the extension.

Press Windows and E keys at the same time to display the Explorer.

Check the extension of the file name from the display.

If this is an image file, .jpeg and .png.

If it is a music file, .wav, .mp3, etc. will be displayed at the end.

You can see the type of file at a glance.

Here are the prerequisites.

If it is a gaming PC purchased within 3 years, I think it is roughly satisfied.

Access the page of Github's Automatic1111.

There is an item called Installation on Windows in the middle of the page.

There are three things you need to install in advance.

One is Python, a programming language.

Click on the link and jump to the download page of Python.

Press and hold the Ctrl key and F key at the same time to display the search dialog on the page.

Enter 3.10.6 in the dialog.

Then you will be able to jump to 3.10.6, so click on the Installer 64-bit at the bottom and download it.

If the version is different, errors may occur, so be careful.

If you can download it, run the file.

Then you will see a screen like this, so check the bottom "Add Python to PATH".

Don't forget to mark it.

This is to remember the location of the installed Python, so you don't have to write a long address during the program.

Even if you don't write a few lines and numbers on the letter, if you write Python, it will be delivered to you.

Click "Install Now" and Success.

Yay!

It's a good opportunity, so let's use the command prompt to check the version of the installed Python.

What is a command prompt?

Press the Windows key and R key at the same time on the desktop and enter CMD there.

It's what Super computer-geek uses.

You can check it by entering "Python space, hyphen, hyphen, and version" there.

If display 3.10.6, you can install it properly.

When you close it, either enter an exit or click the X in the upper right.

This is called a command prompt, and I'll use it again later, so remember it.

And the second thing you need is Git.

This is also from Automatic 1111.

Click the Click Here To Download at the top.

The latest stable version for 64-bit Windows will be downloaded.

If you can download it, run it, and you can press Next.

Once you have installed Python and Git, decide where to install the Stable Diffusion Web UI.

I don't want to install it on C drive with my habit, so this time I'm going to create a new folder right below the D drive and install it.

Open the D drive, right-click, and create a new folder.

I put in Stable Diffusion and Today's Date.

At this time, please do not use Double-Byte-Characters such as Japanese in the folder name.

This is a tradition since time immemorial when installing foreign software.

If the root folder contains DBC such as Japanese, it may fail to install.

Even if it goes well, there may be an error when installing another extensions later.

When installing software, remember that the folder name should be with Single-byte-Characters.

By the way, this install folder will be accessed frequently in the future, so it is recommended to create a shortcut around the desktop.

Once you have a folder to install, click the empty part of the Explorer address bar in that folder.

Then enter CMD.

Now the command prompt has opened in the position of this folder.

It's the one we used to check the version of Python earlier.

Select "Git clone ~ .git" on the page of Automatic1111 and copy it.

Go back to the command prompt, press the Ctrl key and V key to paste.

This means duplicating the file from the page of Automatic1111 in the folder I just created in the D drive.

Press Enter to run.

When the cursor pops up and pops up in the same way as the first time, it's done.

If you go back to the folder, you can see that a lot of files have been generated.

This is the essensial files of Stable Diffusion Web UI Automatic1111.

Actually, we can't use it yet.

It's not written on the page of Automatic1111, but you need to get at least one model.

Models are based on the taste of the generated illustration.

There are more and more models, such as those that are good at anime characters and those that are good at live-action illustrations, so prepare the one that suits your purpose.

Which model do you use?

This time, I would like to install AnythingV4, a standard model of anime illustrations.

Let's get a model!

Search for Hugging Face in the browser and access.

Enter Anything in the search window of Hugging Face.

Then you will see some candidates, so click AnythingV4 by Andite.

If you scroll through the page, you will see a sample, and it seems that you can generate illustrations like this.

Cute!

It seems that you are also good at boys and landscape paintings.

Click "Files and Versions" above.

The model is the one with the extension .checkpoint or .safetensors in the file list that is being released.

The capacity is about the size of a gigabyte file.

There is a reason why it is better to have an empty capacity.

I made it possible to display the extension earlier so that I could identify the dropped file.

This person seems to be uploading a lot of models, and there seems to be Anything V3, V4, and V4.5.

Which one should I choose?

This time, I will choose the pruned and safetensors version from V4.

Many models have a float.16 version, float.32 version, pruned version, safe tensors version, and a full-size version, but to be honest, I don't feel much difference.

Therefore, it seems that the pruned and safetensors version is preferred for the balance of security and capacity.

It would be good to select the one with these labels as a priority.

If you do this, it will be downloaded automatically.

Once you have downloaded it, put the file in the model folder of the web UI that you deployed earlier.

D drive, the folder I made earlier.

Open it because there is a folder called Stable Diffusion Web UI in it.

Open it because there is a folder called Models in it.

Open the folder called Stable Diffusion.

Put checkpoint here!

It's easy to understand because there is a memo saying that the model is OK here.

If it looks like this, it's OK.

This is a folder for storing models, and you will be able to access it again when installing a new model, so keep it in mind.

Launch for the first time!

I'm ready to launch, so I'll launch for the first time.

There is a batch file called "webui-user.bat" in the deployed file.

This is the file to be used when launching.

Double-click to run.

Then a suspicious screen will appear and automatically start preparing to launch.

It takes a while to complete.

It takes time to launch for the first time.

The waiting time depends on the machine performance, but it may take about 10 minutes to 1 hour.

After waiting for a while, if it says "Running on Local URL" "To Create Public Link", the preparation is complete.

Yay!

Don't close this screen, leave it as it is.

Select from "http ~ 7860" by mouse.

Then press the Ctrl key and C key to copy.

Paste it in the URL column of the browser with Ctrl V, press Enter to access it.

Wow!

This is the web UI operating screen.

Let's register this page as you like.

Registering will make it easier to start the next time.

I'll mass-produce the HOT image right away.

Leave it as it is and check the operation first.

I think there is a command called Stable Diffusion Checkpoint on the top left of the screen.

This is the model currently in use.

If you confirm that the model name you just installed is displayed there,

press the orange Generate button on the right.

After a while, an image will be generated.

Something is coming out.

Once the blur is output, the operation check is successful.

The installation is going well and it is in a state where it can be used.

Yay!

Let's HOT images right away!

Now that the operation has been confirmed, I will explain how to end the web UI and how to start it the next time.

When you finish, just close the browser and the prompt screen.

Well, according to theory.

To start, double-click the "webUI-user.bat" file as before.

Wait a while for the prompt screen to start.

After confirming that it came out with Create Public Link in the same way,

call the web UI from browser.

When you start, you need to do it in the order of batch file, standby, and access in the browser.

It's not a must, but as a bonus, I'll put in "Xformers".

After finishing the web UI, right-click the batch file and select Edit from the other options.

Then you can open the batch file in a notepad.

Enter "--Xformers" at the end of the set command line arguments in it.

If you write this down, the image generation will be a little faster.

If you can, save it and finish it.

With this, "Xformers" will be automatically activated when it is activated after the next time.

Let's use it!

I was able to install it safely, so let's actually use it.

First of all, I will briefly explain the interface.

I think you can see that there are several tabs under the model name on the top left.

What you can do with this tab is different.

The default tab is the Text to Image tab.

The number "2" is the English "TO".

In other words, it means to generate images from text and sentences.

In short, T2i is also written.

Next to it is Image to Image, a function that generates images from images.

In short, this is also written as I2i.

The function called Inpaint, which corrects a part, is also included here.

Next is the Extra tab.

It is mainly used for upscaling images.

It is easy to use because you can just specify the magnification method and magnification ratio.

Next is the PNG Info tab.

This is a function that displays the information engraved on the PNG image.

In the image output in the Stable Diffusion Web UI,

information on what kind of settings were output is engraved.

You can view that information and use it again when you re-output it in the Text to Image.

You can also use it to look at the settings of the AI images created by other people.

The following Checkpoint Merger tab and Train tab are for advanced users.

Checkpoint Merger is a synthetic function of the model as its name suggests.

For example, you can combine a model that is good at anime illustrations and a model that is good at live-action styles

to create a model that matches the characteristics of both companies.

I sometimes call the combined model a "Merged model".

The Train tab is a learning function.

You can use it when you want to learn more about your favorite works.

There are ways to learn, such as "Hypernetwork", "Dreambooth", and "LoRA", which is a hot topic recently.

It's a little difficult, so you don't have to worry about it at first.

The setting tab next to it is set as its name suggests.

Extensions are used to add and install expansion functions.

This time, I will explain how to use the basic Text to Image.

In the Text to Image tab, I will introduce the basic operation while actually generating the image.

Since it is an interface at the time of February 23,

it may look a little different when the installation time is different.

First, we will operate from the default state right after launching.

The prompt and negative prompt column in the Text to Image tab.

This is where you write what you want to generate by AI.

I'm going to ask AI to draw a picture like this.

First, I'll type "Girl" in the prompt column above.

It means to draw a girl.

Press the Generate button on the right to wait for the output.

A girl came out.

You can see that it became a girl more clearly than the blur that was generated by the start-up confirmation.

Next, let's make it a smile.

I'll add a "smile" to the prompt column.

Separate each prompt by comma.

When I generated it again,

a picture of a smile was generated.

I think you already understand.

All you have to do is write down the attribute of the illustration you want to be drawn in the prompt column.

It's good to remember that prompt equals to draw everyone's personality.

By the way, the output image is automatically saved in the Outputs folder in the installed folder.

Does this mean that i can draw a tight skirt hot lady or a little loli girl?

Of course you can draw.

For example, let's write it like this.

Change the girl to a woman and add an office lady and a tight skirt.

Then...

Wow!

By the way, why is the image smoked?

Good question.

This is because "VAE" is not applicable.

"United Arab Emirates"??

Think of it as a seasoning that affects the texture of the illustration.

Let's get VAE!

There are two types of VAE.

One is designed by the creator specifically for that model,

and another is a "common VAE".

For example, if it is the current Anything V4,

a small file called "VAE.py.touch" on the page of the download-page on Hugging Face is VAE file.

It's easy to understand because it often has the same name as the model name.

There are several types of VAE called standard,

but this time I will install the one from the Stability AI.

Access Hugging Face again and enter Stability AI in the search window,

and you will see VAE Original in the top right corner, so click on it.

There is a file called VAE 840,000 in the Files and Versions where you click.

This is a VAE file.

It's 840,000, so I'll call it "84VAE".

I will download the SafeTensors version again this time.

Where should I install it?

There are two ways to apply VAE.

There are automatic and manual applications.

However, if possible, it is better to learn how to apply it manually,

so I will explain in that way this time.

The place to put the VAE file is in the Install folder.

There is a folder called Models, and there is a folder called VAE in it.

Put it in this folder.

If the web UI is running, please close and restart browser and batch.

Open the Settings tab after restarting.

Click Stable Diffusion from the column on the left.

There is an item called SD VAE.

I think it's "Automatic" by default.

If you click it, it will pull down, so change it to the 840,000 you put in earlier.

If you set it to 840,000, press the Apply Settings above to apply it.

Will the smoke clear with this?

Let's compare the difference wuth the tight skirt lady.

First, open the PNG Info tab of the web UI.

Open the Outputs folder in the Install folder.

Then, the image you created earlier is saved in it.

The lady from earlier!

Drag and drop this lady to PNG Info.

Then, the prompt, seed value, and model name used for creation will be displayed in this way.

Click Send to T2i.

This will copy the prompt and seed value to the T2i screen.

You re-summon the same lady using the same spell and seed value.

Press the Generate button to generate the image.

Oh, it's a vivid!

The smoke cleared with VAE applied.

Let's compare them side by side.

It's totally different!

As a side note, VAE stands for "Variational Auto Encoder".

In AI illustrations, it has the effect of slightly changing the finish like this comparison.

I think you've learned how to use VAE and PNG Info, so let's continue with T2i.

You're going to teach me how to spell prompts, right?

The way of writing prompts, are different for each person.

If there are 100 users, there are 100 ways to write.

Also, depending on the model being used, the way of writing that is recommended is different,

so if you write it like this, there is no perfect way of writing.

However, in the creation of a 2D character illustration, a certain degree of theory has been decided.

Let's learn how to spell!

First of all, prompts are taken into account as strongly as the one in front, that is, the one on the left.

This is the basics of the basics, so don't forget it!

Based on those characteristics, quality spellings are composed in the first place by style, environment, and main-body.

You can change the order of the three in the back, or you can divide them or mix them.

However, i recommend to keep quality in front.

What is "quality" prompts?

Let's take a look at the tight skirt I mentioned earlier.

As before, it starts with T2i sent from PNG Info.

Quality pts are intended to improve the overall quality, as the name suggests.

As an example, there is such pts.

In the part of "~", there are adjectives such as High, Ultra, and Best.

I'll add Masterpiece to my lady's head.

If you generate it with this,

Face is a little clearer now.

When compared, the impression has changed a lot.

I feel a little goofy.

Next, I'll add "Best Quality, Ultra Detailed".

It looks like an erotic ad.

And these quality-spells are also effective for negative prompts.

Don't you draw elements you don't want to be negative?

That's right.

In other words, if you write a crappy quality-spell that you don't want,

It's a way of thinking that the overall quality will increase.

I'll add "Worst Quality and Low Quality" to the negative column.

It's a fluffy character.

But I feel like I'm getting closer to a pretty girl illustration.

If you compare them, you'll see the importance of quality-spells.

First of all, for the time being,

It's okay to recognize that it's okay if you mark Masterpiece and Quality.

Now that I've drawn the quality, I'm going to add the style and environment.

The style is an illustration, a sketch, a watercolor, a watercolor, etc.

You can also add the name of a famous painter in history to bring the style closer.

The environment includes the whole picture, bust-up, and other composition information,

In addition to the time zone, season, weather, Japanese style, medieval European style, etc.

It includes the background as a whole.

For example, I'll draw it like this.

Add a sketch and a watercolor to the back of the quality.

For the "main-body" part, I'm going to change the woman to One Girl Haz Silver Long Hair and make it silver hair long.

In addition, to emphasize the eyes, add eyelash and jewelry eye, etc.

Look at the viewer's eyes.

Finally, as an environmental spelling, it's outdoor, street, and night.

It's like an after-five of a silver-haired office lady.

The point here is that you can write not only words, but also sentences.

In addition, I'm writing a weather warning spell on the negative prompt.

I'll write it in the summary section of this video.

I'll try to output it with this.

Wow!

I'll compare it with the previous one.

The eyes are beautiful and the amount of information around them has increased.

Next, let's touch on the setting part.

The sampling method on the top left is default, and it's an Euler.

It's fast to generate, but it's a bit of a taste, so I'll change it.

Which one should I use?

It seems that the lower half of the "Karras series" is preferred for illustration generation.

I'll write down most of the features in the summary section.

This time, I'll use the "2M Karras".

This is what happens when you try to change only the method and output it.

We're going to wake up together in tomorrow morning.

If you increase the number of steps on the right, the details will increase, but the generation will be slower.

Moreover, even if you increase it too much, there are many things that don't change much, so it seems that 20 to 40 are preferred.

This time, I'll use 40.

The "Restore face" is the one that corrects the face as its name suggests, but it's a scary face in a 2D illustration, so I won't use it.

The "Tiling" is also not used when drawing person.

"Hi-Resfix" functions as an upscaler.

It highlights what is generated at the resolution of the lower size, taking into account the prompts.

The settings are really dull, so it's for advanced users.

Width and Height are the resolution of the image.

It seems that AI is basically thinking in 512.

If you make it too large, body will split and proning to deformity, so be careful.

If it's a single person, humans are vertically long, so it's good to keep the resolution a little vertical.

Batch counts means how many times generate.

Batch size means how many draw images per single generate.

CFG scale means how much to follow prompts.

Step 40, Vidos 768, Height 960.

Legs are a little weird, but I think you can see the change when you compare it with the first one.

She've re-trained a lot.

If the pts and sampling settings change in the same model, even in the same seed value, it changes this much.

This is what I output by randomly returning only the seed value.

You keep looking for a miracle while changing the spell and sampling settings like this.

One more, one more.

This "generation-swamp" is the interesting part of the AI illustration.

One more, one more.

What's wrong?

I put a spell in it, but it doesn't come out as I thought it would. What should I do?

Oh, you want to generate a flat-chest character, but AI come up with big boob?

In that case, let's use an emphasized spell.

Emphasized spell?

Emphasized spell, that is, an emphasized spellings.

For example, if you want to output a new character, you can put it in a flat chest, but sometimes it doesn't work as you think.

In that case, circle it in brackets like this.

This will determine the flat chest to be 1.1 times.

What did you say?

In addition, the brackets can be layered.

In the case of double brackets, 1.1 times 2 is 1.21 times 2, and 30 is 1.33 times 2.

You can also put multiple words together in brackets.

For example, if you want to emphasize skinny in addition to the flat chest.

In this case, circle it in brackets like this.

This will emphasize both words by 1.1 times.

And I don't think there are many opportunities to use it, but if you circle it in brackets, it will be weak against it.

It is 0.91 times 0.83 times.

This can be used if you want to add just a little element in terms of hidden taste.

And if it's a hassle to hit a lot of brackets, you can also use the method of specifying by numbers.

For example, if you want to emphasize the flat chest by 1.3 times,

In the case of multiple cases, if you circle it in brackets with 1.3 in the end,

All the elements in the brackets will be emphasized.

In addition, if you find it troublesome to type,

After tracing the words you want to emphasize with the mouse,

You can set the strength and weakness with the up and down arrows while pressing the control key.

Let's take a look at the changes.

For example, let's emphasize the older sister on board again.

It's a status.

I'll try to sandwich it with a flat chest.

Then, I'll try to double the flat chest by 1.1.

I see. Your face and body are a little out of sync.

In that case, let's make the girl a little girl.

Wow! It's a magical girl outside the office!

It's important to match other elements.

And what I want you to pay attention to is that

Please remember that the emphasis is about 1.5 times more important.

Also, depending on the model, it can cause the picture to break down.

If you feel that the picture is getting dirty, you can hold it down or scrape it.

It's important to arrange everything.

If you've seen the video so far, I think you can already use it in a basic way.

In that case, I'd like to try another model.

So, at the end, I'll introduce you to some of my favorite models.

First of all, "Anything V3".

It is a standard model of the benchmark existence of a beautiful girl model.

In this video, I took up V4 as an example.

However, the texture is different, so it is difficult to press and hold.

"ACertainthing"

It is an image model of V3 with more background information.

Personally, when I asked for a sketch prompt,

Outline is clearer than V3.

It's an easy-to-make model.

"Seventh Anime V3"

There are multiple versions with different textures.

This is a model called V3C.

It's delicate, but it's got a solid outline.

And the sides are narrow, so it's easy to see the changes in VAE when changing.

"Abyss Orange Mix 3"

It's a high-quality model with a higher resolution of the face.

By default, it's a pale texture.

If you like a beautiful girl illustration like a doll, you might like it.

"Counterfeit V2.5"

It's a delicate touch of a beautiful girl model.

It's a model with features that don't have anything in the outline.

It's a model of the impression that foreigners think of Moe Bishoujo.

"Pastel Mix"

This is what the same person as Anything V4 and V4.5 is publishing.

It's a model with a characteristic touch like a pastel painting.

Anyway, it's a model with primary colors.

The scenery is beautiful like a painting.

It's a level that can be printed out and displayed.

"Basil Mix"

It's a live-action model specialized for Asian models.

It's a model that can do a lot of things.

It's a model that's a little difficult to handle to output beautifully, such as adjusting the resolution to fit the human head and body.

Only a small part of the models featured in this video.

The AI illustration is making great progress.

At this moment, new models and new expansion features are being released all over the world.

That's all for today.

See you next week!